?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Use auxv to check for CRC32 instructions on ARM.
[pgsql.git] / src / backend / executor / nodeHashjoin.c
blob6c3009fba0fc6ac4518c343a1425642c756655c3
1 /*-------------------------------------------------------------------------
2 *
3 * nodeHashjoin.c
4 * Routines to handle hash join nodes
5 *
6 * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
7 * Portions Copyright (c) 1994, Regents of the University of California
8 *
9 *
10 * IDENTIFICATION
11 * src/backend/executor/nodeHashjoin.c
12 *
13 * HASH JOIN
14 *
15 * This is based on the "hybrid hash join" algorithm described shortly in the
16 * following page
17 *
18 * https://en.wikipedia.org/wiki/Hash_join#Hybrid_hash_join
19 *
20 * and in detail in the referenced paper:
21 *
22 * "An Adaptive Hash Join Algorithm for Multiuser Environments"
23 * Hansjörg Zeller; Jim Gray (1990). Proceedings of the 16th VLDB conference.
24 * Brisbane: 186–197.
25 *
26 * If the inner side tuples of a hash join do not fit in memory, the hash join
27 * can be executed in multiple batches.
28 *
29 * If the statistics on the inner side relation are accurate, planner chooses a
30 * multi-batch strategy and estimates the number of batches.
31 *
32 * The query executor measures the real size of the hashtable and increases the
33 * number of batches if the hashtable grows too large.
34 *
35 * The number of batches is always a power of two, so an increase in the number
36 * of batches doubles it.
37 *
38 * Serial hash join measures batch size lazily -- waiting until it is loading a
39 * batch to determine if it will fit in memory. While inserting tuples into the
40 * hashtable, serial hash join will, if that tuple were to exceed work_mem,
41 * dump out the hashtable and reassign them either to other batch files or the
42 * current batch resident in the hashtable.
43 *
44 * Parallel hash join, on the other hand, completes all changes to the number
45 * of batches during the build phase. If it increases the number of batches, it
46 * dumps out all the tuples from all batches and reassigns them to entirely new
47 * batch files. Then it checks every batch to ensure it will fit in the space
48 * budget for the query.
49 *
50 * In both parallel and serial hash join, the executor currently makes a best
51 * effort. If a particular batch will not fit in memory, it tries doubling the
52 * number of batches. If after a batch increase, there is a batch which
53 * retained all or none of its tuples, the executor disables growth in the
54 * number of batches globally. After growth is disabled, all batches that would
55 * have previously triggered an increase in the number of batches instead
56 * exceed the space allowed.
57 *
58 * PARALLELISM
59 *
60 * Hash joins can participate in parallel query execution in several ways. A
61 * parallel-oblivious hash join is one where the node is unaware that it is
62 * part of a parallel plan. In this case, a copy of the inner plan is used to
63 * build a copy of the hash table in every backend, and the outer plan could
64 * either be built from a partial or complete path, so that the results of the
65 * hash join are correspondingly either partial or complete. A parallel-aware
66 * hash join is one that behaves differently, coordinating work between
67 * backends, and appears as Parallel Hash Join in EXPLAIN output. A Parallel
68 * Hash Join always appears with a Parallel Hash node.
69 *
70 * Parallel-aware hash joins use the same per-backend state machine to track
71 * progress through the hash join algorithm as parallel-oblivious hash joins.
72 * In a parallel-aware hash join, there is also a shared state machine that
73 * co-operating backends use to synchronize their local state machines and
74 * program counters. The shared state machine is managed with a Barrier IPC
75 * primitive. When all attached participants arrive at a barrier, the phase
76 * advances and all waiting participants are released.
77 *
78 * When a participant begins working on a parallel hash join, it must first
79 * figure out how much progress has already been made, because participants
80 * don't wait for each other to begin. For this reason there are switch
81 * statements at key points in the code where we have to synchronize our local
82 * state machine with the phase, and then jump to the correct part of the
83 * algorithm so that we can get started.
84 *
85 * One barrier called build_barrier is used to coordinate the hashing phases.
86 * The phase is represented by an integer which begins at zero and increments
87 * one by one, but in the code it is referred to by symbolic names as follows.
88 * An asterisk indicates a phase that is performed by a single arbitrarily
89 * chosen process.
90 *
91 * PHJ_BUILD_ELECT -- initial state
92 * PHJ_BUILD_ALLOCATE* -- one sets up the batches and table 0
93 * PHJ_BUILD_HASH_INNER -- all hash the inner rel
94 * PHJ_BUILD_HASH_OUTER -- (multi-batch only) all hash the outer
95 * PHJ_BUILD_RUN -- building done, probing can begin
96 * PHJ_BUILD_FREE* -- all work complete, one frees batches
97 *
98 * While in the phase PHJ_BUILD_HASH_INNER a separate pair of barriers may
99 * be used repeatedly as required to coordinate expansions in the number of
100 * batches or buckets. Their phases are as follows:
101 *
102 * PHJ_GROW_BATCHES_ELECT -- initial state
103 * PHJ_GROW_BATCHES_REALLOCATE* -- one allocates new batches
104 * PHJ_GROW_BATCHES_REPARTITION -- all repartition
105 * PHJ_GROW_BATCHES_DECIDE* -- one detects skew and cleans up
106 * PHJ_GROW_BATCHES_FINISH -- finished one growth cycle
107 *
108 * PHJ_GROW_BUCKETS_ELECT -- initial state
109 * PHJ_GROW_BUCKETS_REALLOCATE* -- one allocates new buckets
110 * PHJ_GROW_BUCKETS_REINSERT -- all insert tuples
111 *
112 * If the planner got the number of batches and buckets right, those won't be
113 * necessary, but on the other hand we might finish up needing to expand the
114 * buckets or batches multiple times while hashing the inner relation to stay
115 * within our memory budget and load factor target. For that reason it's a
116 * separate pair of barriers using circular phases.
117 *
118 * The PHJ_BUILD_HASH_OUTER phase is required only for multi-batch joins,
119 * because we need to divide the outer relation into batches up front in order
120 * to be able to process batches entirely independently. In contrast, the
121 * parallel-oblivious algorithm simply throws tuples 'forward' to 'later'
122 * batches whenever it encounters them while scanning and probing, which it
123 * can do because it processes batches in serial order.
124 *
125 * Once PHJ_BUILD_RUN is reached, backends then split up and process
126 * different batches, or gang up and work together on probing batches if there
127 * aren't enough to go around. For each batch there is a separate barrier
128 * with the following phases:
129 *
130 * PHJ_BATCH_ELECT -- initial state
131 * PHJ_BATCH_ALLOCATE* -- one allocates buckets
132 * PHJ_BATCH_LOAD -- all load the hash table from disk
133 * PHJ_BATCH_PROBE -- all probe
134 * PHJ_BATCH_SCAN* -- one does right/right-anti/full unmatched scan
135 * PHJ_BATCH_FREE* -- one frees memory
136 *
137 * Batch 0 is a special case, because it starts out in phase
138 * PHJ_BATCH_PROBE; populating batch 0's hash table is done during
139 * PHJ_BUILD_HASH_INNER so we can skip loading.
140 *
141 * Initially we try to plan for a single-batch hash join using the combined
142 * hash_mem of all participants to create a large shared hash table. If that
143 * turns out either at planning or execution time to be impossible then we
144 * fall back to regular hash_mem sized hash tables.
145 *
146 * To avoid deadlocks, we never wait for any barrier unless it is known that
147 * all other backends attached to it are actively executing the node or have
148 * finished. Practically, that means that we never emit a tuple while attached
149 * to a barrier, unless the barrier has reached a phase that means that no
150 * process will wait on it again. We emit tuples while attached to the build
151 * barrier in phase PHJ_BUILD_RUN, and to a per-batch barrier in phase
152 * PHJ_BATCH_PROBE. These are advanced to PHJ_BUILD_FREE and PHJ_BATCH_SCAN
153 * respectively without waiting, using BarrierArriveAndDetach() and
154 * BarrierArriveAndDetachExceptLast() respectively. The last to detach
155 * receives a different return value so that it knows that it's safe to
156 * clean up. Any straggler process that attaches after that phase is reached
157 * will see that it's too late to participate or access the relevant shared
158 * memory objects.
159 *
160 *-------------------------------------------------------------------------
161 */
162
163 #include "postgres.h"
164
165 #include "access/htup_details.h"
166 #include "access/parallel.h"
167 #include "executor/executor.h"
168 #include "executor/hashjoin.h"
169 #include "executor/nodeHash.h"
170 #include "executor/nodeHashjoin.h"
171 #include "miscadmin.h"
172 #include "utils/lsyscache.h"
173 #include "utils/sharedtuplestore.h"
174 #include "utils/wait_event.h"
175
176
177 /*
178 * States of the ExecHashJoin state machine
179 */
180 #define HJ_BUILD_HASHTABLE 1
181 #define HJ_NEED_NEW_OUTER 2
182 #define HJ_SCAN_BUCKET 3
183 #define HJ_FILL_OUTER_TUPLE 4
184 #define HJ_FILL_INNER_TUPLES 5
185 #define HJ_NEED_NEW_BATCH 6
186
187 /* Returns true if doing null-fill on outer relation */
188 #define HJ_FILL_OUTER(hjstate) ((hjstate)->hj_NullInnerTupleSlot != NULL)
189 /* Returns true if doing null-fill on inner relation */
190 #define HJ_FILL_INNER(hjstate) ((hjstate)->hj_NullOuterTupleSlot != NULL)
191
192 static TupleTableSlot *ExecHashJoinOuterGetTuple(PlanState *outerNode,
193 HashJoinState *hjstate,
194 uint32 *hashvalue);
195 static TupleTableSlot *ExecParallelHashJoinOuterGetTuple(PlanState *outerNode,
196 HashJoinState *hjstate,
197 uint32 *hashvalue);
198 static TupleTableSlot *ExecHashJoinGetSavedTuple(HashJoinState *hjstate,
199 BufFile *file,
200 uint32 *hashvalue,
201 TupleTableSlot *tupleSlot);
202 static bool ExecHashJoinNewBatch(HashJoinState *hjstate);
203 static bool ExecParallelHashJoinNewBatch(HashJoinState *hjstate);
204 static void ExecParallelHashJoinPartitionOuter(HashJoinState *hjstate);
205
206
207 /* ----------------------------------------------------------------
208 * ExecHashJoinImpl
209 *
210 * This function implements the Hybrid Hashjoin algorithm. It is marked
211 * with an always-inline attribute so that ExecHashJoin() and
212 * ExecParallelHashJoin() can inline it. Compilers that respect the
213 * attribute should create versions specialized for parallel == true and
214 * parallel == false with unnecessary branches removed.
215 *
216 * Note: the relation we build hash table on is the "inner"
217 * the other one is "outer".
218 * ----------------------------------------------------------------
219 */
220 static pg_attribute_always_inline TupleTableSlot *
221 ExecHashJoinImpl(PlanState *pstate, bool parallel)
222 {
223 HashJoinState *node = castNode(HashJoinState, pstate);
224 PlanState *outerNode;
225 HashState *hashNode;
226 ExprState *joinqual;
227 ExprState *otherqual;
228 ExprContext *econtext;
229 HashJoinTable hashtable;
230 TupleTableSlot *outerTupleSlot;
231 uint32 hashvalue;
232 int batchno;
233 ParallelHashJoinState *parallel_state;
234
235 /*
236 * get information from HashJoin node
237 */
238 joinqual = node->js.joinqual;
239 otherqual = node->js.ps.qual;
240 hashNode = (HashState *) innerPlanState(node);
241 outerNode = outerPlanState(node);
242 hashtable = node->hj_HashTable;
243 econtext = node->js.ps.ps_ExprContext;
244 parallel_state = hashNode->parallel_state;
245
246 /*
247 * Reset per-tuple memory context to free any expression evaluation
248 * storage allocated in the previous tuple cycle.
249 */
250 ResetExprContext(econtext);
251
252 /*
253 * run the hash join state machine
254 */
255 for (;;)
256 {
257 /*
258 * It's possible to iterate this loop many times before returning a
259 * tuple, in some pathological cases such as needing to move much of
260 * the current batch to a later batch. So let's check for interrupts
261 * each time through.
262 */
263 CHECK_FOR_INTERRUPTS();
264
265 switch (node->hj_JoinState)
266 {
267 case HJ_BUILD_HASHTABLE:
268
269 /*
270 * First time through: build hash table for inner relation.
271 */
272 Assert(hashtable == NULL);
273
274 /*
275 * If the outer relation is completely empty, and it's not
276 * right/right-anti/full join, we can quit without building
277 * the hash table. However, for an inner join it is only a
278 * win to check this when the outer relation's startup cost is
279 * less than the projected cost of building the hash table.
280 * Otherwise it's best to build the hash table first and see
281 * if the inner relation is empty. (When it's a left join, we
282 * should always make this check, since we aren't going to be
283 * able to skip the join on the strength of an empty inner
284 * relation anyway.)
285 *
286 * If we are rescanning the join, we make use of information
287 * gained on the previous scan: don't bother to try the
288 * prefetch if the previous scan found the outer relation
289 * nonempty. This is not 100% reliable since with new
290 * parameters the outer relation might yield different
291 * results, but it's a good heuristic.
292 *
293 * The only way to make the check is to try to fetch a tuple
294 * from the outer plan node. If we succeed, we have to stash
295 * it away for later consumption by ExecHashJoinOuterGetTuple.
296 */
297 if (HJ_FILL_INNER(node))
298 {
299 /* no chance to not build the hash table */
300 node->hj_FirstOuterTupleSlot = NULL;
301 }
302 else if (parallel)
303 {
304 /*
305 * The empty-outer optimization is not implemented for
306 * shared hash tables, because no one participant can
307 * determine that there are no outer tuples, and it's not
308 * yet clear that it's worth the synchronization overhead
309 * of reaching consensus to figure that out. So we have
310 * to build the hash table.
311 */
312 node->hj_FirstOuterTupleSlot = NULL;
313 }
314 else if (HJ_FILL_OUTER(node) ||
315 (outerNode->plan->startup_cost < hashNode->ps.plan->total_cost &&
316 !node->hj_OuterNotEmpty))
317 {
318 node->hj_FirstOuterTupleSlot = ExecProcNode(outerNode);
319 if (TupIsNull(node->hj_FirstOuterTupleSlot))
320 {
321 node->hj_OuterNotEmpty = false;
322 return NULL;
323 }
324 else
325 node->hj_OuterNotEmpty = true;
326 }
327 else
328 node->hj_FirstOuterTupleSlot = NULL;
329
330 /*
331 * Create the hash table. If using Parallel Hash, then
332 * whoever gets here first will create the hash table and any
333 * later arrivals will merely attach to it.
334 */
335 hashtable = ExecHashTableCreate(hashNode);
336 node->hj_HashTable = hashtable;
337
338 /*
339 * Execute the Hash node, to build the hash table. If using
340 * Parallel Hash, then we'll try to help hashing unless we
341 * arrived too late.
342 */
343 hashNode->hashtable = hashtable;
344 (void) MultiExecProcNode((PlanState *) hashNode);
345
346 /*
347 * If the inner relation is completely empty, and we're not
348 * doing a left outer join, we can quit without scanning the
349 * outer relation.
350 */
351 if (hashtable->totalTuples == 0 && !HJ_FILL_OUTER(node))
352 {
353 if (parallel)
354 {
355 /*
356 * Advance the build barrier to PHJ_BUILD_RUN before
357 * proceeding so we can negotiate resource cleanup.
358 */
359 Barrier *build_barrier = &parallel_state->build_barrier;
360
361 while (BarrierPhase(build_barrier) < PHJ_BUILD_RUN)
362 BarrierArriveAndWait(build_barrier, 0);
363 }
364 return NULL;
365 }
366
367 /*
368 * need to remember whether nbatch has increased since we
369 * began scanning the outer relation
370 */
371 hashtable->nbatch_outstart = hashtable->nbatch;
372
373 /*
374 * Reset OuterNotEmpty for scan. (It's OK if we fetched a
375 * tuple above, because ExecHashJoinOuterGetTuple will
376 * immediately set it again.)
377 */
378 node->hj_OuterNotEmpty = false;
379
380 if (parallel)
381 {
382 Barrier *build_barrier;
383
384 build_barrier = &parallel_state->build_barrier;
385 Assert(BarrierPhase(build_barrier) == PHJ_BUILD_HASH_OUTER ||
386 BarrierPhase(build_barrier) == PHJ_BUILD_RUN ||
387 BarrierPhase(build_barrier) == PHJ_BUILD_FREE);
388 if (BarrierPhase(build_barrier) == PHJ_BUILD_HASH_OUTER)
389 {
390 /*
391 * If multi-batch, we need to hash the outer relation
392 * up front.
393 */
394 if (hashtable->nbatch > 1)
395 ExecParallelHashJoinPartitionOuter(node);
396 BarrierArriveAndWait(build_barrier,
397 WAIT_EVENT_HASH_BUILD_HASH_OUTER);
398 }
399 else if (BarrierPhase(build_barrier) == PHJ_BUILD_FREE)
400 {
401 /*
402 * If we attached so late that the job is finished and
403 * the batch state has been freed, we can return
404 * immediately.
405 */
406 return NULL;
407 }
408
409 /* Each backend should now select a batch to work on. */
410 Assert(BarrierPhase(build_barrier) == PHJ_BUILD_RUN);
411 hashtable->curbatch = -1;
412 node->hj_JoinState = HJ_NEED_NEW_BATCH;
413
414 continue;
415 }
416 else
417 node->hj_JoinState = HJ_NEED_NEW_OUTER;
418
419 /* FALL THRU */
420
421 case HJ_NEED_NEW_OUTER:
422
423 /*
424 * We don't have an outer tuple, try to get the next one
425 */
426 if (parallel)
427 outerTupleSlot =
428 ExecParallelHashJoinOuterGetTuple(outerNode, node,
429 &hashvalue);
430 else
431 outerTupleSlot =
432 ExecHashJoinOuterGetTuple(outerNode, node, &hashvalue);
433
434 if (TupIsNull(outerTupleSlot))
435 {
436 /* end of batch, or maybe whole join */
437 if (HJ_FILL_INNER(node))
438 {
439 /* set up to scan for unmatched inner tuples */
440 if (parallel)
441 {
442 /*
443 * Only one process is currently allow to handle
444 * each batch's unmatched tuples, in a parallel
445 * join.
446 */
447 if (ExecParallelPrepHashTableForUnmatched(node))
448 node->hj_JoinState = HJ_FILL_INNER_TUPLES;
449 else
450 node->hj_JoinState = HJ_NEED_NEW_BATCH;
451 }
452 else
453 {
454 ExecPrepHashTableForUnmatched(node);
455 node->hj_JoinState = HJ_FILL_INNER_TUPLES;
456 }
457 }
458 else
459 node->hj_JoinState = HJ_NEED_NEW_BATCH;
460 continue;
461 }
462
463 econtext->ecxt_outertuple = outerTupleSlot;
464 node->hj_MatchedOuter = false;
465
466 /*
467 * Find the corresponding bucket for this tuple in the main
468 * hash table or skew hash table.
469 */
470 node->hj_CurHashValue = hashvalue;
471 ExecHashGetBucketAndBatch(hashtable, hashvalue,
472 &node->hj_CurBucketNo, &batchno);
473 node->hj_CurSkewBucketNo = ExecHashGetSkewBucket(hashtable,
474 hashvalue);
475 node->hj_CurTuple = NULL;
476
477 /*
478 * The tuple might not belong to the current batch (where
479 * "current batch" includes the skew buckets if any).
480 */
481 if (batchno != hashtable->curbatch &&
482 node->hj_CurSkewBucketNo == INVALID_SKEW_BUCKET_NO)
483 {
484 bool shouldFree;
485 MinimalTuple mintuple = ExecFetchSlotMinimalTuple(outerTupleSlot,
486 &shouldFree);
487
488 /*
489 * Need to postpone this outer tuple to a later batch.
490 * Save it in the corresponding outer-batch file.
491 */
492 Assert(parallel_state == NULL);
493 Assert(batchno > hashtable->curbatch);
494 ExecHashJoinSaveTuple(mintuple, hashvalue,
495 &hashtable->outerBatchFile[batchno],
496 hashtable);
497
498 if (shouldFree)
499 heap_free_minimal_tuple(mintuple);
500
501 /* Loop around, staying in HJ_NEED_NEW_OUTER state */
502 continue;
503 }
504
505 /* OK, let's scan the bucket for matches */
506 node->hj_JoinState = HJ_SCAN_BUCKET;
507
508 /* FALL THRU */
509
510 case HJ_SCAN_BUCKET:
511
512 /*
513 * Scan the selected hash bucket for matches to current outer
514 */
515 if (parallel)
516 {
517 if (!ExecParallelScanHashBucket(node, econtext))
518 {
519 /* out of matches; check for possible outer-join fill */
520 node->hj_JoinState = HJ_FILL_OUTER_TUPLE;
521 continue;
522 }
523 }
524 else
525 {
526 if (!ExecScanHashBucket(node, econtext))
527 {
528 /* out of matches; check for possible outer-join fill */
529 node->hj_JoinState = HJ_FILL_OUTER_TUPLE;
530 continue;
531 }
532 }
533
534 /*
535 * In a right-semijoin, we only need the first match for each
536 * inner tuple.
537 */
538 if (node->js.jointype == JOIN_RIGHT_SEMI &&
539 HeapTupleHeaderHasMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple)))
540 continue;
541
542 /*
543 * We've got a match, but still need to test non-hashed quals.
544 * ExecScanHashBucket already set up all the state needed to
545 * call ExecQual.
546 *
547 * If we pass the qual, then save state for next call and have
548 * ExecProject form the projection, store it in the tuple
549 * table, and return the slot.
550 *
551 * Only the joinquals determine tuple match status, but all
552 * quals must pass to actually return the tuple.
553 */
554 if (joinqual == NULL || ExecQual(joinqual, econtext))
555 {
556 node->hj_MatchedOuter = true;
557
558 /*
559 * This is really only needed if HJ_FILL_INNER(node) or if
560 * we are in a right-semijoin, but we'll avoid the branch
561 * and just set it always.
562 */
563 if (!HeapTupleHeaderHasMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple)))
564 HeapTupleHeaderSetMatch(HJTUPLE_MINTUPLE(node->hj_CurTuple));
565
566 /* In an antijoin, we never return a matched tuple */
567 if (node->js.jointype == JOIN_ANTI)
568 {
569 node->hj_JoinState = HJ_NEED_NEW_OUTER;
570 continue;
571 }
572
573 /*
574 * If we only need to consider the first matching inner
575 * tuple, then advance to next outer tuple after we've
576 * processed this one.
577 */
578 if (node->js.single_match)
579 node->hj_JoinState = HJ_NEED_NEW_OUTER;
580
581 /*
582 * In a right-antijoin, we never return a matched tuple.
583 * If it's not an inner_unique join, we need to stay on
584 * the current outer tuple to continue scanning the inner
585 * side for matches.
586 */
587 if (node->js.jointype == JOIN_RIGHT_ANTI)
588 continue;
589
590 if (otherqual == NULL || ExecQual(otherqual, econtext))
591 return ExecProject(node->js.ps.ps_ProjInfo);
592 else
593 InstrCountFiltered2(node, 1);
594 }
595 else
596 InstrCountFiltered1(node, 1);
597 break;
598
599 case HJ_FILL_OUTER_TUPLE:
600
601 /*
602 * The current outer tuple has run out of matches, so check
603 * whether to emit a dummy outer-join tuple. Whether we emit
604 * one or not, the next state is NEED_NEW_OUTER.
605 */
606 node->hj_JoinState = HJ_NEED_NEW_OUTER;
607
608 if (!node->hj_MatchedOuter &&
609 HJ_FILL_OUTER(node))
610 {
611 /*
612 * Generate a fake join tuple with nulls for the inner
613 * tuple, and return it if it passes the non-join quals.
614 */
615 econtext->ecxt_innertuple = node->hj_NullInnerTupleSlot;
616
617 if (otherqual == NULL || ExecQual(otherqual, econtext))
618 return ExecProject(node->js.ps.ps_ProjInfo);
619 else
620 InstrCountFiltered2(node, 1);
621 }
622 break;
623
624 case HJ_FILL_INNER_TUPLES:
625
626 /*
627 * We have finished a batch, but we are doing
628 * right/right-anti/full join, so any unmatched inner tuples
629 * in the hashtable have to be emitted before we continue to
630 * the next batch.
631 */
632 if (!(parallel ? ExecParallelScanHashTableForUnmatched(node, econtext)
633 : ExecScanHashTableForUnmatched(node, econtext)))
634 {
635 /* no more unmatched tuples */
636 node->hj_JoinState = HJ_NEED_NEW_BATCH;
637 continue;
638 }
639
640 /*
641 * Generate a fake join tuple with nulls for the outer tuple,
642 * and return it if it passes the non-join quals.
643 */
644 econtext->ecxt_outertuple = node->hj_NullOuterTupleSlot;
645
646 if (otherqual == NULL || ExecQual(otherqual, econtext))
647 return ExecProject(node->js.ps.ps_ProjInfo);
648 else
649 InstrCountFiltered2(node, 1);
650 break;
651
652 case HJ_NEED_NEW_BATCH:
653
654 /*
655 * Try to advance to next batch. Done if there are no more.
656 */
657 if (parallel)
658 {
659 if (!ExecParallelHashJoinNewBatch(node))
660 return NULL; /* end of parallel-aware join */
661 }
662 else
663 {
664 if (!ExecHashJoinNewBatch(node))
665 return NULL; /* end of parallel-oblivious join */
666 }
667 node->hj_JoinState = HJ_NEED_NEW_OUTER;
668 break;
669
670 default:
671 elog(ERROR, "unrecognized hashjoin state: %d",
672 (int) node->hj_JoinState);
673 }
674 }
675 }
676
677 /* ----------------------------------------------------------------
678 * ExecHashJoin
679 *
680 * Parallel-oblivious version.
681 * ----------------------------------------------------------------
682 */
683 static TupleTableSlot * /* return: a tuple or NULL */
684 ExecHashJoin(PlanState *pstate)
685 {
686 /*
687 * On sufficiently smart compilers this should be inlined with the
688 * parallel-aware branches removed.
689 */
690 return ExecHashJoinImpl(pstate, false);
691 }
692
693 /* ----------------------------------------------------------------
694 * ExecParallelHashJoin
695 *
696 * Parallel-aware version.
697 * ----------------------------------------------------------------
698 */
699 static TupleTableSlot * /* return: a tuple or NULL */
700 ExecParallelHashJoin(PlanState *pstate)
701 {
702 /*
703 * On sufficiently smart compilers this should be inlined with the
704 * parallel-oblivious branches removed.
705 */
706 return ExecHashJoinImpl(pstate, true);
707 }
708
709 /* ----------------------------------------------------------------
710 * ExecInitHashJoin
711 *
712 * Init routine for HashJoin node.
713 * ----------------------------------------------------------------
714 */
715 HashJoinState *
716 ExecInitHashJoin(HashJoin *node, EState *estate, int eflags)
717 {
718 HashJoinState *hjstate;
719 Plan *outerNode;
720 Hash *hashNode;
721 TupleDesc outerDesc,
722 innerDesc;
723 const TupleTableSlotOps *ops;
724
725 /* check for unsupported flags */
726 Assert(!(eflags & (EXEC_FLAG_BACKWARD | EXEC_FLAG_MARK)));
727
728 /*
729 * create state structure
730 */
731 hjstate = makeNode(HashJoinState);
732 hjstate->js.ps.plan = (Plan *) node;
733 hjstate->js.ps.state = estate;
734
735 /*
736 * See ExecHashJoinInitializeDSM() and ExecHashJoinInitializeWorker()
737 * where this function may be replaced with a parallel version, if we
738 * managed to launch a parallel query.
739 */
740 hjstate->js.ps.ExecProcNode = ExecHashJoin;
741 hjstate->js.jointype = node->join.jointype;
742
743 /*
744 * Miscellaneous initialization
745 *
746 * create expression context for node
747 */
748 ExecAssignExprContext(estate, &hjstate->js.ps);
749
750 /*
751 * initialize child nodes
752 *
753 * Note: we could suppress the REWIND flag for the inner input, which
754 * would amount to betting that the hash will be a single batch. Not
755 * clear if this would be a win or not.
756 */
757 outerNode = outerPlan(node);
758 hashNode = (Hash *) innerPlan(node);
759
760 outerPlanState(hjstate) = ExecInitNode(outerNode, estate, eflags);
761 outerDesc = ExecGetResultType(outerPlanState(hjstate));
762 innerPlanState(hjstate) = ExecInitNode((Plan *) hashNode, estate, eflags);
763 innerDesc = ExecGetResultType(innerPlanState(hjstate));
764
765 /*
766 * Initialize result slot, type and projection.
767 */
768 ExecInitResultTupleSlotTL(&hjstate->js.ps, &TTSOpsVirtual);
769 ExecAssignProjectionInfo(&hjstate->js.ps, NULL);
770
771 /*
772 * tuple table initialization
773 */
774 ops = ExecGetResultSlotOps(outerPlanState(hjstate), NULL);
775 hjstate->hj_OuterTupleSlot = ExecInitExtraTupleSlot(estate, outerDesc,
776 ops);
777
778 /*
779 * detect whether we need only consider the first matching inner tuple
780 */
781 hjstate->js.single_match = (node->join.inner_unique ||
782 node->join.jointype == JOIN_SEMI);
783
784 /* set up null tuples for outer joins, if needed */
785 switch (node->join.jointype)
786 {
787 case JOIN_INNER:
788 case JOIN_SEMI:
789 case JOIN_RIGHT_SEMI:
790 break;
791 case JOIN_LEFT:
792 case JOIN_ANTI:
793 hjstate->hj_NullInnerTupleSlot =
794 ExecInitNullTupleSlot(estate, innerDesc, &TTSOpsVirtual);
795 break;
796 case JOIN_RIGHT:
797 case JOIN_RIGHT_ANTI:
798 hjstate->hj_NullOuterTupleSlot =
799 ExecInitNullTupleSlot(estate, outerDesc, &TTSOpsVirtual);
800 break;
801 case JOIN_FULL:
802 hjstate->hj_NullOuterTupleSlot =
803 ExecInitNullTupleSlot(estate, outerDesc, &TTSOpsVirtual);
804 hjstate->hj_NullInnerTupleSlot =
805 ExecInitNullTupleSlot(estate, innerDesc, &TTSOpsVirtual);
806 break;
807 default:
808 elog(ERROR, "unrecognized join type: %d",
809 (int) node->join.jointype);
810 }
811
812 /*
813 * now for some voodoo. our temporary tuple slot is actually the result
814 * tuple slot of the Hash node (which is our inner plan). we can do this
815 * because Hash nodes don't return tuples via ExecProcNode() -- instead
816 * the hash join node uses ExecScanHashBucket() to get at the contents of
817 * the hash table. -cim 6/9/91
818 */
819 {
820 HashState *hashstate = (HashState *) innerPlanState(hjstate);
821 Hash *hash = (Hash *) hashstate->ps.plan;
822 TupleTableSlot *slot = hashstate->ps.ps_ResultTupleSlot;
823 Oid *outer_hashfuncid;
824 Oid *inner_hashfuncid;
825 bool *hash_strict;
826 ListCell *lc;
827 int nkeys;
828
829
830 hjstate->hj_HashTupleSlot = slot;
831
832 /*
833 * Build ExprStates to obtain hash values for either side of the join.
834 * This must be done here as ExecBuildHash32Expr needs to know how to
835 * handle NULL inputs and the required handling of that depends on the
836 * jointype. We don't know the join type in ExecInitHash() and we
837 * must build the ExprStates before ExecHashTableCreate() so we
838 * properly attribute any SubPlans that exist in the hash expressions
839 * to the correct PlanState.
840 */
841 nkeys = list_length(node->hashoperators);
842
843 outer_hashfuncid = palloc_array(Oid, nkeys);
844 inner_hashfuncid = palloc_array(Oid, nkeys);
845 hash_strict = palloc_array(bool, nkeys);
846
847 /*
848 * Determine the hash function for each side of the join for the given
849 * hash operator.
850 */
851 foreach(lc, node->hashoperators)
852 {
853 Oid hashop = lfirst_oid(lc);
854 int i = foreach_current_index(lc);
855
856 if (!get_op_hash_functions(hashop,
857 &outer_hashfuncid[i],
858 &inner_hashfuncid[i]))
859 elog(ERROR,
860 "could not find hash function for hash operator %u",
861 hashop);
862 hash_strict[i] = op_strict(hashop);
863 }
864
865 /*
866 * Build an ExprState to generate the hash value for the expressions
867 * on the outer of the join. This ExprState must finish generating
868 * the hash value when HJ_FILL_OUTER() is true. Otherwise,
869 * ExecBuildHash32Expr will set up the ExprState to abort early if it
870 * finds a NULL. In these cases, we don't need to store these tuples
871 * in the hash table as the jointype does not require it.
872 */
873 hjstate->hj_OuterHash =
874 ExecBuildHash32Expr(hjstate->js.ps.ps_ResultTupleDesc,
875 hjstate->js.ps.resultops,
876 outer_hashfuncid,
877 node->hashcollations,
878 node->hashkeys,
879 hash_strict,
880 &hjstate->js.ps,
881 0,
882 HJ_FILL_OUTER(hjstate));
883
884 /* As above, but for the inner side of the join */
885 hashstate->hash_expr =
886 ExecBuildHash32Expr(hashstate->ps.ps_ResultTupleDesc,
887 hashstate->ps.resultops,
888 inner_hashfuncid,
889 node->hashcollations,
890 hash->hashkeys,
891 hash_strict,
892 &hashstate->ps,
893 0,
894 HJ_FILL_INNER(hjstate));
895
896 /*
897 * Set up the skew table hash function while we have a record of the
898 * first key's hash function Oid.
899 */
900 if (OidIsValid(hash->skewTable))
901 {
902 hashstate->skew_hashfunction = palloc0(sizeof(FmgrInfo));
903 hashstate->skew_collation = linitial_oid(node->hashcollations);
904 fmgr_info(outer_hashfuncid[0], hashstate->skew_hashfunction);
905 }
906
907 /* no need to keep these */
908 pfree(outer_hashfuncid);
909 pfree(inner_hashfuncid);
910 pfree(hash_strict);
911 }
912
913 /*
914 * initialize child expressions
915 */
916 hjstate->js.ps.qual =
917 ExecInitQual(node->join.plan.qual, (PlanState *) hjstate);
918 hjstate->js.joinqual =
919 ExecInitQual(node->join.joinqual, (PlanState *) hjstate);
920 hjstate->hashclauses =
921 ExecInitQual(node->hashclauses, (PlanState *) hjstate);
922
923 /*
924 * initialize hash-specific info
925 */
926 hjstate->hj_HashTable = NULL;
927 hjstate->hj_FirstOuterTupleSlot = NULL;
928
929 hjstate->hj_CurHashValue = 0;
930 hjstate->hj_CurBucketNo = 0;
931 hjstate->hj_CurSkewBucketNo = INVALID_SKEW_BUCKET_NO;
932 hjstate->hj_CurTuple = NULL;
933
934 hjstate->hj_JoinState = HJ_BUILD_HASHTABLE;
935 hjstate->hj_MatchedOuter = false;
936 hjstate->hj_OuterNotEmpty = false;
937
938 return hjstate;
939 }
940
941 /* ----------------------------------------------------------------
942 * ExecEndHashJoin
943 *
944 * clean up routine for HashJoin node
945 * ----------------------------------------------------------------
946 */
947 void
948 ExecEndHashJoin(HashJoinState *node)
949 {
950 /*
951 * Free hash table
952 */
953 if (node->hj_HashTable)
954 {
955 ExecHashTableDestroy(node->hj_HashTable);
956 node->hj_HashTable = NULL;
957 }
958
959 /*
960 * clean up subtrees
961 */
962 ExecEndNode(outerPlanState(node));
963 ExecEndNode(innerPlanState(node));
964 }
965
966 /*
967 * ExecHashJoinOuterGetTuple
968 *
969 * get the next outer tuple for a parallel oblivious hashjoin: either by
970 * executing the outer plan node in the first pass, or from the temp
971 * files for the hashjoin batches.
972 *
973 * Returns a null slot if no more outer tuples (within the current batch).
974 *
975 * On success, the tuple's hash value is stored at *hashvalue --- this is
976 * either originally computed, or re-read from the temp file.
977 */
978 static TupleTableSlot *
979 ExecHashJoinOuterGetTuple(PlanState *outerNode,
980 HashJoinState *hjstate,
981 uint32 *hashvalue)
982 {
983 HashJoinTable hashtable = hjstate->hj_HashTable;
984 int curbatch = hashtable->curbatch;
985 TupleTableSlot *slot;
986
987 if (curbatch == 0) /* if it is the first pass */
988 {
989 /*
990 * Check to see if first outer tuple was already fetched by
991 * ExecHashJoin() and not used yet.
992 */
993 slot = hjstate->hj_FirstOuterTupleSlot;
994 if (!TupIsNull(slot))
995 hjstate->hj_FirstOuterTupleSlot = NULL;
996 else
997 slot = ExecProcNode(outerNode);
998
999 while (!TupIsNull(slot))
1000 {
1001 bool isnull;
1002
1003 /*
1004 * We have to compute the tuple's hash value.
1005 */
1006 ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
1007
1008 econtext->ecxt_outertuple = slot;
1009
1010 ResetExprContext(econtext);
1011
1012 *hashvalue = DatumGetUInt32(ExecEvalExprSwitchContext(hjstate->hj_OuterHash,
1013 econtext,
1014 &isnull));
1015
1016 if (!isnull)
1017 {
1018 /* remember outer relation is not empty for possible rescan */
1019 hjstate->hj_OuterNotEmpty = true;
1020
1021 return slot;
1022 }
1023
1024 /*
1025 * That tuple couldn't match because of a NULL, so discard it and
1026 * continue with the next one.
1027 */
1028 slot = ExecProcNode(outerNode);
1029 }
1030 }
1031 else if (curbatch < hashtable->nbatch)
1032 {
1033 BufFile *file = hashtable->outerBatchFile[curbatch];
1034
1035 /*
1036 * In outer-join cases, we could get here even though the batch file
1037 * is empty.
1038 */
1039 if (file == NULL)
1040 return NULL;
1041
1042 slot = ExecHashJoinGetSavedTuple(hjstate,
1043 file,
1044 hashvalue,
1045 hjstate->hj_OuterTupleSlot);
1046 if (!TupIsNull(slot))
1047 return slot;
1048 }
1049
1050 /* End of this batch */
1051 return NULL;
1052 }
1053
1054 /*
1055 * ExecHashJoinOuterGetTuple variant for the parallel case.
1056 */
1057 static TupleTableSlot *
1058 ExecParallelHashJoinOuterGetTuple(PlanState *outerNode,
1059 HashJoinState *hjstate,
1060 uint32 *hashvalue)
1061 {
1062 HashJoinTable hashtable = hjstate->hj_HashTable;
1063 int curbatch = hashtable->curbatch;
1064 TupleTableSlot *slot;
1065
1066 /*
1067 * In the Parallel Hash case we only run the outer plan directly for
1068 * single-batch hash joins. Otherwise we have to go to batch files, even
1069 * for batch 0.
1070 */
1071 if (curbatch == 0 && hashtable->nbatch == 1)
1072 {
1073 slot = ExecProcNode(outerNode);
1074
1075 while (!TupIsNull(slot))
1076 {
1077 bool isnull;
1078
1079 ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
1080
1081 econtext->ecxt_outertuple = slot;
1082
1083 ResetExprContext(econtext);
1084
1085 *hashvalue = DatumGetUInt32(ExecEvalExprSwitchContext(hjstate->hj_OuterHash,
1086 econtext,
1087 &isnull));
1088
1089 if (!isnull)
1090 return slot;
1091
1092 /*
1093 * That tuple couldn't match because of a NULL, so discard it and
1094 * continue with the next one.
1095 */
1096 slot = ExecProcNode(outerNode);
1097 }
1098 }
1099 else if (curbatch < hashtable->nbatch)
1100 {
1101 MinimalTuple tuple;
1102
1103 tuple = sts_parallel_scan_next(hashtable->batches[curbatch].outer_tuples,
1104 hashvalue);
1105 if (tuple != NULL)
1106 {
1107 ExecForceStoreMinimalTuple(tuple,
1108 hjstate->hj_OuterTupleSlot,
1109 false);
1110 slot = hjstate->hj_OuterTupleSlot;
1111 return slot;
1112 }
1113 else
1114 ExecClearTuple(hjstate->hj_OuterTupleSlot);
1115 }
1116
1117 /* End of this batch */
1118 hashtable->batches[curbatch].outer_eof = true;
1119
1120 return NULL;
1121 }
1122
1123 /*
1124 * ExecHashJoinNewBatch
1125 * switch to a new hashjoin batch
1126 *
1127 * Returns true if successful, false if there are no more batches.
1128 */
1129 static bool
1130 ExecHashJoinNewBatch(HashJoinState *hjstate)
1131 {
1132 HashJoinTable hashtable = hjstate->hj_HashTable;
1133 int nbatch;
1134 int curbatch;
1135 BufFile *innerFile;
1136 TupleTableSlot *slot;
1137 uint32 hashvalue;
1138
1139 nbatch = hashtable->nbatch;
1140 curbatch = hashtable->curbatch;
1141
1142 if (curbatch > 0)
1143 {
1144 /*
1145 * We no longer need the previous outer batch file; close it right
1146 * away to free disk space.
1147 */
1148 if (hashtable->outerBatchFile[curbatch])
1149 BufFileClose(hashtable->outerBatchFile[curbatch]);
1150 hashtable->outerBatchFile[curbatch] = NULL;
1151 }
1152 else /* we just finished the first batch */
1153 {
1154 /*
1155 * Reset some of the skew optimization state variables, since we no
1156 * longer need to consider skew tuples after the first batch. The
1157 * memory context reset we are about to do will release the skew
1158 * hashtable itself.
1159 */
1160 hashtable->skewEnabled = false;
1161 hashtable->skewBucket = NULL;
1162 hashtable->skewBucketNums = NULL;
1163 hashtable->nSkewBuckets = 0;
1164 hashtable->spaceUsedSkew = 0;
1165 }
1166
1167 /*
1168 * We can always skip over any batches that are completely empty on both
1169 * sides. We can sometimes skip over batches that are empty on only one
1170 * side, but there are exceptions:
1171 *
1172 * 1. In a left/full outer join, we have to process outer batches even if
1173 * the inner batch is empty. Similarly, in a right/right-anti/full outer
1174 * join, we have to process inner batches even if the outer batch is
1175 * empty.
1176 *
1177 * 2. If we have increased nbatch since the initial estimate, we have to
1178 * scan inner batches since they might contain tuples that need to be
1179 * reassigned to later inner batches.
1180 *
1181 * 3. Similarly, if we have increased nbatch since starting the outer
1182 * scan, we have to rescan outer batches in case they contain tuples that
1183 * need to be reassigned.
1184 */
1185 curbatch++;
1186 while (curbatch < nbatch &&
1187 (hashtable->outerBatchFile[curbatch] == NULL ||
1188 hashtable->innerBatchFile[curbatch] == NULL))
1189 {
1190 if (hashtable->outerBatchFile[curbatch] &&
1191 HJ_FILL_OUTER(hjstate))
1192 break; /* must process due to rule 1 */
1193 if (hashtable->innerBatchFile[curbatch] &&
1194 HJ_FILL_INNER(hjstate))
1195 break; /* must process due to rule 1 */
1196 if (hashtable->innerBatchFile[curbatch] &&
1197 nbatch != hashtable->nbatch_original)
1198 break; /* must process due to rule 2 */
1199 if (hashtable->outerBatchFile[curbatch] &&
1200 nbatch != hashtable->nbatch_outstart)
1201 break; /* must process due to rule 3 */
1202 /* We can ignore this batch. */
1203 /* Release associated temp files right away. */
1204 if (hashtable->innerBatchFile[curbatch])
1205 BufFileClose(hashtable->innerBatchFile[curbatch]);
1206 hashtable->innerBatchFile[curbatch] = NULL;
1207 if (hashtable->outerBatchFile[curbatch])
1208 BufFileClose(hashtable->outerBatchFile[curbatch]);
1209 hashtable->outerBatchFile[curbatch] = NULL;
1210 curbatch++;
1211 }
1212
1213 if (curbatch >= nbatch)
1214 return false; /* no more batches */
1215
1216 hashtable->curbatch = curbatch;
1217
1218 /*
1219 * Reload the hash table with the new inner batch (which could be empty)
1220 */
1221 ExecHashTableReset(hashtable);
1222
1223 innerFile = hashtable->innerBatchFile[curbatch];
1224
1225 if (innerFile != NULL)
1226 {
1227 if (BufFileSeek(innerFile, 0, 0, SEEK_SET))
1228 ereport(ERROR,
1229 (errcode_for_file_access(),
1230 errmsg("could not rewind hash-join temporary file")));
1231
1232 while ((slot = ExecHashJoinGetSavedTuple(hjstate,
1233 innerFile,
1234 &hashvalue,
1235 hjstate->hj_HashTupleSlot)))
1236 {
1237 /*
1238 * NOTE: some tuples may be sent to future batches. Also, it is
1239 * possible for hashtable->nbatch to be increased here!
1240 */
1241 ExecHashTableInsert(hashtable, slot, hashvalue);
1242 }
1243
1244 /*
1245 * after we build the hash table, the inner batch file is no longer
1246 * needed
1247 */
1248 BufFileClose(innerFile);
1249 hashtable->innerBatchFile[curbatch] = NULL;
1250 }
1251
1252 /*
1253 * Rewind outer batch file (if present), so that we can start reading it.
1254 */
1255 if (hashtable->outerBatchFile[curbatch] != NULL)
1256 {
1257 if (BufFileSeek(hashtable->outerBatchFile[curbatch], 0, 0, SEEK_SET))
1258 ereport(ERROR,
1259 (errcode_for_file_access(),
1260 errmsg("could not rewind hash-join temporary file")));
1261 }
1262
1263 return true;
1264 }
1265
1266 /*
1267 * Choose a batch to work on, and attach to it. Returns true if successful,
1268 * false if there are no more batches.
1269 */
1270 static bool
1271 ExecParallelHashJoinNewBatch(HashJoinState *hjstate)
1272 {
1273 HashJoinTable hashtable = hjstate->hj_HashTable;
1274 int start_batchno;
1275 int batchno;
1276
1277 /*
1278 * If we were already attached to a batch, remember not to bother checking
1279 * it again, and detach from it (possibly freeing the hash table if we are
1280 * last to detach).
1281 */
1282 if (hashtable->curbatch >= 0)
1283 {
1284 hashtable->batches[hashtable->curbatch].done = true;
1285 ExecHashTableDetachBatch(hashtable);
1286 }
1287
1288 /*
1289 * Search for a batch that isn't done. We use an atomic counter to start
1290 * our search at a different batch in every participant when there are
1291 * more batches than participants.
1292 */
1293 batchno = start_batchno =
1294 pg_atomic_fetch_add_u32(&hashtable->parallel_state->distributor, 1) %
1295 hashtable->nbatch;
1296 do
1297 {
1298 uint32 hashvalue;
1299 MinimalTuple tuple;
1300 TupleTableSlot *slot;
1301
1302 if (!hashtable->batches[batchno].done)
1303 {
1304 SharedTuplestoreAccessor *inner_tuples;
1305 Barrier *batch_barrier =
1306 &hashtable->batches[batchno].shared->batch_barrier;
1307
1308 switch (BarrierAttach(batch_barrier))
1309 {
1310 case PHJ_BATCH_ELECT:
1311
1312 /* One backend allocates the hash table. */
1313 if (BarrierArriveAndWait(batch_barrier,
1314 WAIT_EVENT_HASH_BATCH_ELECT))
1315 ExecParallelHashTableAlloc(hashtable, batchno);
1316 /* Fall through. */
1317
1318 case PHJ_BATCH_ALLOCATE:
1319 /* Wait for allocation to complete. */
1320 BarrierArriveAndWait(batch_barrier,
1321 WAIT_EVENT_HASH_BATCH_ALLOCATE);
1322 /* Fall through. */
1323
1324 case PHJ_BATCH_LOAD:
1325 /* Start (or join in) loading tuples. */
1326 ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1327 inner_tuples = hashtable->batches[batchno].inner_tuples;
1328 sts_begin_parallel_scan(inner_tuples);
1329 while ((tuple = sts_parallel_scan_next(inner_tuples,
1330 &hashvalue)))
1331 {
1332 ExecForceStoreMinimalTuple(tuple,
1333 hjstate->hj_HashTupleSlot,
1334 false);
1335 slot = hjstate->hj_HashTupleSlot;
1336 ExecParallelHashTableInsertCurrentBatch(hashtable, slot,
1337 hashvalue);
1338 }
1339 sts_end_parallel_scan(inner_tuples);
1340 BarrierArriveAndWait(batch_barrier,
1341 WAIT_EVENT_HASH_BATCH_LOAD);
1342 /* Fall through. */
1343
1344 case PHJ_BATCH_PROBE:
1345
1346 /*
1347 * This batch is ready to probe. Return control to
1348 * caller. We stay attached to batch_barrier so that the
1349 * hash table stays alive until everyone's finished
1350 * probing it, but no participant is allowed to wait at
1351 * this barrier again (or else a deadlock could occur).
1352 * All attached participants must eventually detach from
1353 * the barrier and one worker must advance the phase so
1354 * that the final phase is reached.
1355 */
1356 ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1357 sts_begin_parallel_scan(hashtable->batches[batchno].outer_tuples);
1358
1359 return true;
1360 case PHJ_BATCH_SCAN:
1361
1362 /*
1363 * In principle, we could help scan for unmatched tuples,
1364 * since that phase is already underway (the thing we
1365 * can't do under current deadlock-avoidance rules is wait
1366 * for others to arrive at PHJ_BATCH_SCAN, because
1367 * PHJ_BATCH_PROBE emits tuples, but in this case we just
1368 * got here without waiting). That is not yet done. For
1369 * now, we just detach and go around again. We have to
1370 * use ExecHashTableDetachBatch() because there's a small
1371 * chance we'll be the last to detach, and then we're
1372 * responsible for freeing memory.
1373 */
1374 ExecParallelHashTableSetCurrentBatch(hashtable, batchno);
1375 hashtable->batches[batchno].done = true;
1376 ExecHashTableDetachBatch(hashtable);
1377 break;
1378
1379 case PHJ_BATCH_FREE:
1380
1381 /*
1382 * Already done. Detach and go around again (if any
1383 * remain).
1384 */
1385 BarrierDetach(batch_barrier);
1386 hashtable->batches[batchno].done = true;
1387 hashtable->curbatch = -1;
1388 break;
1389
1390 default:
1391 elog(ERROR, "unexpected batch phase %d",
1392 BarrierPhase(batch_barrier));
1393 }
1394 }
1395 batchno = (batchno + 1) % hashtable->nbatch;
1396 } while (batchno != start_batchno);
1397
1398 return false;
1399 }
1400
1401 /*
1402 * ExecHashJoinSaveTuple
1403 * save a tuple to a batch file.
1404 *
1405 * The data recorded in the file for each tuple is its hash value,
1406 * then the tuple in MinimalTuple format.
1407 *
1408 * fileptr points to a batch file in one of the hashtable arrays.
1409 *
1410 * The batch files (and their buffers) are allocated in the spill context
1411 * created for the hashtable.
1412 */
1413 void
1414 ExecHashJoinSaveTuple(MinimalTuple tuple, uint32 hashvalue,
1415 BufFile **fileptr, HashJoinTable hashtable)
1416 {
1417 BufFile *file = *fileptr;
1418
1419 /*
1420 * The batch file is lazily created. If this is the first tuple written to
1421 * this batch, the batch file is created and its buffer is allocated in
1422 * the spillCxt context, NOT in the batchCxt.
1423 *
1424 * During the build phase, buffered files are created for inner batches.
1425 * Each batch's buffered file is closed (and its buffer freed) after the
1426 * batch is loaded into memory during the outer side scan. Therefore, it
1427 * is necessary to allocate the batch file buffer in a memory context
1428 * which outlives the batch itself.
1429 *
1430 * Also, we use spillCxt instead of hashCxt for a better accounting of the
1431 * spilling memory consumption.
1432 */
1433 if (file == NULL)
1434 {
1435 MemoryContext oldctx = MemoryContextSwitchTo(hashtable->spillCxt);
1436
1437 file = BufFileCreateTemp(false);
1438 *fileptr = file;
1439
1440 MemoryContextSwitchTo(oldctx);
1441 }
1442
1443 BufFileWrite(file, &hashvalue, sizeof(uint32));
1444 BufFileWrite(file, tuple, tuple->t_len);
1445 }
1446
1447 /*
1448 * ExecHashJoinGetSavedTuple
1449 * read the next tuple from a batch file. Return NULL if no more.
1450 *
1451 * On success, *hashvalue is set to the tuple's hash value, and the tuple
1452 * itself is stored in the given slot.
1453 */
1454 static TupleTableSlot *
1455 ExecHashJoinGetSavedTuple(HashJoinState *hjstate,
1456 BufFile *file,
1457 uint32 *hashvalue,
1458 TupleTableSlot *tupleSlot)
1459 {
1460 uint32 header[2];
1461 size_t nread;
1462 MinimalTuple tuple;
1463
1464 /*
1465 * We check for interrupts here because this is typically taken as an
1466 * alternative code path to an ExecProcNode() call, which would include
1467 * such a check.
1468 */
1469 CHECK_FOR_INTERRUPTS();
1470
1471 /*
1472 * Since both the hash value and the MinimalTuple length word are uint32,
1473 * we can read them both in one BufFileRead() call without any type
1474 * cheating.
1475 */
1476 nread = BufFileReadMaybeEOF(file, header, sizeof(header), true);
1477 if (nread == 0) /* end of file */
1478 {
1479 ExecClearTuple(tupleSlot);
1480 return NULL;
1481 }
1482 *hashvalue = header[0];
1483 tuple = (MinimalTuple) palloc(header[1]);
1484 tuple->t_len = header[1];
1485 BufFileReadExact(file,
1486 (char *) tuple + sizeof(uint32),
1487 header[1] - sizeof(uint32));
1488 ExecForceStoreMinimalTuple(tuple, tupleSlot, true);
1489 return tupleSlot;
1490 }
1491
1492
1493 void
1494 ExecReScanHashJoin(HashJoinState *node)
1495 {
1496 PlanState *outerPlan = outerPlanState(node);
1497 PlanState *innerPlan = innerPlanState(node);
1498
1499 /*
1500 * In a multi-batch join, we currently have to do rescans the hard way,
1501 * primarily because batch temp files may have already been released. But
1502 * if it's a single-batch join, and there is no parameter change for the
1503 * inner subnode, then we can just re-use the existing hash table without
1504 * rebuilding it.
1505 */
1506 if (node->hj_HashTable != NULL)
1507 {
1508 if (node->hj_HashTable->nbatch == 1 &&
1509 innerPlan->chgParam == NULL)
1510 {
1511 /*
1512 * Okay to reuse the hash table; needn't rescan inner, either.
1513 *
1514 * However, if it's a right/right-anti/full join, we'd better
1515 * reset the inner-tuple match flags contained in the table.
1516 */
1517 if (HJ_FILL_INNER(node))
1518 ExecHashTableResetMatchFlags(node->hj_HashTable);
1519
1520 /*
1521 * Also, we need to reset our state about the emptiness of the
1522 * outer relation, so that the new scan of the outer will update
1523 * it correctly if it turns out to be empty this time. (There's no
1524 * harm in clearing it now because ExecHashJoin won't need the
1525 * info. In the other cases, where the hash table doesn't exist
1526 * or we are destroying it, we leave this state alone because
1527 * ExecHashJoin will need it the first time through.)
1528 */
1529 node->hj_OuterNotEmpty = false;
1530
1531 /* ExecHashJoin can skip the BUILD_HASHTABLE step */
1532 node->hj_JoinState = HJ_NEED_NEW_OUTER;
1533 }
1534 else
1535 {
1536 /* must destroy and rebuild hash table */
1537 HashState *hashNode = castNode(HashState, innerPlan);
1538
1539 Assert(hashNode->hashtable == node->hj_HashTable);
1540 /* accumulate stats from old hash table, if wanted */
1541 /* (this should match ExecShutdownHash) */
1542 if (hashNode->ps.instrument && !hashNode->hinstrument)
1543 hashNode->hinstrument = (HashInstrumentation *)
1544 palloc0(sizeof(HashInstrumentation));
1545 if (hashNode->hinstrument)
1546 ExecHashAccumInstrumentation(hashNode->hinstrument,
1547 hashNode->hashtable);
1548 /* for safety, be sure to clear child plan node's pointer too */
1549 hashNode->hashtable = NULL;
1550
1551 ExecHashTableDestroy(node->hj_HashTable);
1552 node->hj_HashTable = NULL;
1553 node->hj_JoinState = HJ_BUILD_HASHTABLE;
1554
1555 /*
1556 * if chgParam of subnode is not null then plan will be re-scanned
1557 * by first ExecProcNode.
1558 */
1559 if (innerPlan->chgParam == NULL)
1560 ExecReScan(innerPlan);
1561 }
1562 }
1563
1564 /* Always reset intra-tuple state */
1565 node->hj_CurHashValue = 0;
1566 node->hj_CurBucketNo = 0;
1567 node->hj_CurSkewBucketNo = INVALID_SKEW_BUCKET_NO;
1568 node->hj_CurTuple = NULL;
1569
1570 node->hj_MatchedOuter = false;
1571 node->hj_FirstOuterTupleSlot = NULL;
1572
1573 /*
1574 * if chgParam of subnode is not null then plan will be re-scanned by
1575 * first ExecProcNode.
1576 */
1577 if (outerPlan->chgParam == NULL)
1578 ExecReScan(outerPlan);
1579 }
1580
1581 void
1582 ExecShutdownHashJoin(HashJoinState *node)
1583 {
1584 if (node->hj_HashTable)
1585 {
1586 /*
1587 * Detach from shared state before DSM memory goes away. This makes
1588 * sure that we don't have any pointers into DSM memory by the time
1589 * ExecEndHashJoin runs.
1590 */
1591 ExecHashTableDetachBatch(node->hj_HashTable);
1592 ExecHashTableDetach(node->hj_HashTable);
1593 }
1594 }
1595
1596 static void
1597 ExecParallelHashJoinPartitionOuter(HashJoinState *hjstate)
1598 {
1599 PlanState *outerState = outerPlanState(hjstate);
1600 ExprContext *econtext = hjstate->js.ps.ps_ExprContext;
1601 HashJoinTable hashtable = hjstate->hj_HashTable;
1602 TupleTableSlot *slot;
1603 uint32 hashvalue;
1604 int i;
1605
1606 Assert(hjstate->hj_FirstOuterTupleSlot == NULL);
1607
1608 /* Execute outer plan, writing all tuples to shared tuplestores. */
1609 for (;;)
1610 {
1611 bool isnull;
1612
1613 slot = ExecProcNode(outerState);
1614 if (TupIsNull(slot))
1615 break;
1616 econtext->ecxt_outertuple = slot;
1617
1618 ResetExprContext(econtext);
1619
1620 hashvalue = DatumGetUInt32(ExecEvalExprSwitchContext(hjstate->hj_OuterHash,
1621 econtext,
1622 &isnull));
1623
1624 if (!isnull)
1625 {
1626 int batchno;
1627 int bucketno;
1628 bool shouldFree;
1629 MinimalTuple mintup = ExecFetchSlotMinimalTuple(slot, &shouldFree);
1630
1631 ExecHashGetBucketAndBatch(hashtable, hashvalue, &bucketno,
1632 &batchno);
1633 sts_puttuple(hashtable->batches[batchno].outer_tuples,
1634 &hashvalue, mintup);
1635
1636 if (shouldFree)
1637 heap_free_minimal_tuple(mintup);
1638 }
1639 CHECK_FOR_INTERRUPTS();
1640 }
1641
1642 /* Make sure all outer partitions are readable by any backend. */
1643 for (i = 0; i < hashtable->nbatch; ++i)
1644 sts_end_write(hashtable->batches[i].outer_tuples);
1645 }
1646
1647 void
1648 ExecHashJoinEstimate(HashJoinState *state, ParallelContext *pcxt)
1649 {
1650 shm_toc_estimate_chunk(&pcxt->estimator, sizeof(ParallelHashJoinState));
1651 shm_toc_estimate_keys(&pcxt->estimator, 1);
1652 }
1653
1654 void
1655 ExecHashJoinInitializeDSM(HashJoinState *state, ParallelContext *pcxt)
1656 {
1657 int plan_node_id = state->js.ps.plan->plan_node_id;
1658 HashState *hashNode;
1659 ParallelHashJoinState *pstate;
1660
1661 /*
1662 * Disable shared hash table mode if we failed to create a real DSM
1663 * segment, because that means that we don't have a DSA area to work with.
1664 */
1665 if (pcxt->seg == NULL)
1666 return;
1667
1668 ExecSetExecProcNode(&state->js.ps, ExecParallelHashJoin);
1669
1670 /*
1671 * Set up the state needed to coordinate access to the shared hash
1672 * table(s), using the plan node ID as the toc key.
1673 */
1674 pstate = shm_toc_allocate(pcxt->toc, sizeof(ParallelHashJoinState));
1675 shm_toc_insert(pcxt->toc, plan_node_id, pstate);
1676
1677 /*
1678 * Set up the shared hash join state with no batches initially.
1679 * ExecHashTableCreate() will prepare at least one later and set nbatch
1680 * and space_allowed.
1681 */
1682 pstate->nbatch = 0;
1683 pstate->space_allowed = 0;
1684 pstate->batches = InvalidDsaPointer;
1685 pstate->old_batches = InvalidDsaPointer;
1686 pstate->nbuckets = 0;
1687 pstate->growth = PHJ_GROWTH_OK;
1688 pstate->chunk_work_queue = InvalidDsaPointer;
1689 pg_atomic_init_u32(&pstate->distributor, 0);
1690 pstate->nparticipants = pcxt->nworkers + 1;
1691 pstate->total_tuples = 0;
1692 LWLockInitialize(&pstate->lock,
1693 LWTRANCHE_PARALLEL_HASH_JOIN);
1694 BarrierInit(&pstate->build_barrier, 0);
1695 BarrierInit(&pstate->grow_batches_barrier, 0);
1696 BarrierInit(&pstate->grow_buckets_barrier, 0);
1697
1698 /* Set up the space we'll use for shared temporary files. */
1699 SharedFileSetInit(&pstate->fileset, pcxt->seg);
1700
1701 /* Initialize the shared state in the hash node. */
1702 hashNode = (HashState *) innerPlanState(state);
1703 hashNode->parallel_state = pstate;
1704 }
1705
1706 /* ----------------------------------------------------------------
1707 * ExecHashJoinReInitializeDSM
1708 *
1709 * Reset shared state before beginning a fresh scan.
1710 * ----------------------------------------------------------------
1711 */
1712 void
1713 ExecHashJoinReInitializeDSM(HashJoinState *state, ParallelContext *pcxt)
1714 {
1715 int plan_node_id = state->js.ps.plan->plan_node_id;
1716 ParallelHashJoinState *pstate;
1717
1718 /* Nothing to do if we failed to create a DSM segment. */
1719 if (pcxt->seg == NULL)
1720 return;
1721
1722 pstate = shm_toc_lookup(pcxt->toc, plan_node_id, false);
1723
1724 /*
1725 * It would be possible to reuse the shared hash table in single-batch
1726 * cases by resetting and then fast-forwarding build_barrier to
1727 * PHJ_BUILD_FREE and batch 0's batch_barrier to PHJ_BATCH_PROBE, but
1728 * currently shared hash tables are already freed by now (by the last
1729 * participant to detach from the batch). We could consider keeping it
1730 * around for single-batch joins. We'd also need to adjust
1731 * finalize_plan() so that it doesn't record a dummy dependency for
1732 * Parallel Hash nodes, preventing the rescan optimization. For now we
1733 * don't try.
1734 */
1735
1736 /* Detach, freeing any remaining shared memory. */
1737 if (state->hj_HashTable != NULL)
1738 {
1739 ExecHashTableDetachBatch(state->hj_HashTable);
1740 ExecHashTableDetach(state->hj_HashTable);
1741 }
1742
1743 /* Clear any shared batch files. */
1744 SharedFileSetDeleteAll(&pstate->fileset);
1745
1746 /* Reset build_barrier to PHJ_BUILD_ELECT so we can go around again. */
1747 BarrierInit(&pstate->build_barrier, 0);
1748 }
1749
1750 void
1751 ExecHashJoinInitializeWorker(HashJoinState *state,
1752 ParallelWorkerContext *pwcxt)
1753 {
1754 HashState *hashNode;
1755 int plan_node_id = state->js.ps.plan->plan_node_id;
1756 ParallelHashJoinState *pstate =
1757 shm_toc_lookup(pwcxt->toc, plan_node_id, false);
1758
1759 /* Attach to the space for shared temporary files. */
1760 SharedFileSetAttach(&pstate->fileset, pwcxt->seg);
1761
1762 /* Attach to the shared state in the hash node. */
1763 hashNode = (HashState *) innerPlanState(state);
1764 hashNode->parallel_state = pstate;
1765
1766 ExecSetExecProcNode(&state->js.ps, ExecParallelHashJoin);
1767 }
Postgresql Clone from git://git.postgresql.org/git/postgresql.git